Oldhand: A Portable 'Definition of Done' Workflow for Coding Agents

Coding agents are powerful, but they often lack a crucial ingredient: a clear definition of done. Without it, they can veer off course, implement the wrong thing, or declare victory prematurely. Oldhand, a new open-source project, aims to fix that by providing a portable workflow that wraps around your favorite agent, enforcing a structured process from issue to verified outcome.

Oldhand is a plugin for both Claude Code and ChatGPT Codex. It accepts a ticket from Jira, Asana, monday.com, Linear, or GitHub—or even just a plain prompt—and then guides the agent through a disciplined sequence: understanding the real flow, researching permissive open-source prior art, implementing the smallest safe change (using a helper named Ponytail), and verifying the result end to end through the user's browser or computer-use tools.

The installation is delightfully simple. For Claude Code, you run two prompts to install the plugin, then start a new session. For Codex, you restart or start a new task and use a similar command. There's no need for a tracker integration or MCP server—the plugin contains everything it needs. This portability is a key feature: it works out of the box with just the agent and your existing tools.

A ticket key is optional. You can give Oldhand an implement instruction, but the real power comes from including assumptions, decisions, changed requirements, constraints, partial work, exclusions, relevant files, and the proof you expect. Oldhand treats your latest explicit context as authoritative, reconciles it with older ticket text, and only asks for clarification when an unresolved choice would be unsafe or materially change the result.

What does this actually buy you? Without Oldhand, coding agents often: implement the first plausible solution without checking prior art, skip verification, or make overly broad changes. Oldhand makes the agent: research existing solutions, make minimal changes, and verify end to end before calling the task done. It also adds a layer of accountability—the final verifier is an orchestrator (Fable for Claude, Sol for Codex) that can delegate subtasks to smaller models but always does the final review itself.

It's important to note that Oldhand doesn't grant authority by itself. Remote writes, commits, pushes, pull requests, deployments—these still require your explicit request. The skill just structures the work; you retain control. And if the end-to-end gate is blocked, it's reported as blocked, never as passed.

The project is MIT-licensed and designed to be dependency-free, so you can validate it locally without installing anything beyond the agent itself. For a developer, Oldhand is a pragmatic way to bring more rigor to AI-assisted coding, especially on complex tasks where scope creep is a real risk.

If you're tired of fighting with agents that go off the rails, Oldhand is worth a look. It's a small, clever piece of tooling that could make a big difference in the quality and reliability of agent-generated code.